A statistical model for annotating videos with human actions

This contribution addresses the approach to recognize single and multiple human actions in video streams. This work introduces a novel action recognition algorithm with normalization enhancements. Initially feature vectors are extracted using 2D SIFT features. Bag of Words model is extended with a new normalization technique on the visual vocabulary to make the dimensions same so that the actions would be easier to read and extract. This normalization technique vastly improves the results from the state of the art methods. HMM based model is developed for training and testing of six basic actions present in the KTH human action dataset. By comparing our work with previously applied models, results display that our approach vastly improves the accuracy of the existing methods of action recognition.